Consumer Debt Is Changing Retail Buying Behavior. Here Is What That Does to the Demand Signal an AI Forecasting Tool Was Trained On.

The Monday morning the report looked fine

In August 2026 I was running four things at once: ramping up several AI agent projects for clients while winding down another engagement, trying to stay present in my communities, and setting an intention almost every morning to pause and actually evaluate where my energy was going. The pressure of that much simultaneous movement creates a specific cognitive state. Everything feels urgent, so the things that require patience rather than action get deferred. The intentions that needed a quiet five minutes kept running into a calendar that had no quiet five minutes in it. I kept almost doing the thing I had told myself I would do. That gap between the intention and the action is the place I work from, and it is the place this piece is about. The leaders I work with live inside a version of that same dynamic, and it shows up most visibly on Monday morning. It is 8:47 AM. The weekly operations report lands in the inbox. There are three minutes before the first call, and the summary line matches what was expected last week, so the report gets forwarded to the VP with 'looks good.' On Friday, they had written a note to themselves: ask the analyst to walk through the sourcing on the supplier reconciliation column. That column has been off by a rounding error for six weeks. The note is still in the notebook. The call notification appears. The intention disappears. The line worth carrying out of this piece, the one worth saying to yourself in that three-minute window, is this: a summary line matching your expectation tells you the report closed, and nothing more about whether the numbers are right. That is the sentence for 8:47 AM. Everything below is the evidence behind it.

TL;DR

Four data points and two structural claims that frame everything below.

• Debt at historic levels

  US household debt reached $18.8 trillion in Q4 2025, up $502 billion during the full year, per the Federal Reserve Bank of New York (2025). Credit card balances alone hit $1.28 trillion, a 5.5% annual increase.

• The savings rate collapsed

  The personal savings rate fell to 4.0% in Q1 2026, down from 6.2% in early 2024 (Yahoo Finance / Federal Reserve G.19, 2026). Consumers are spending income they have not yet earned.

• Trade-down is structural, not cyclical

  79% of consumers globally report trading down, with more than half seeking deals on every purchase (McKinsey, as cited in MNTN Research, 2025). In a Trax survey of over 7,000 US shoppers in summer 2025, 66% of those who changed habits switched to cheaper or generic brands.

• The demand signal has split by income tier

  Moody's Analytics estimates that the top 10% of US households accounted for nearly half of all consumer spending in Q2 2025 (Forbes, 2025). A model trained on aggregate demand patterns before that bifurcation will systematically over-forecast mid-market SKUs.

• Model drift is the under-named risk

  A forecasting model trained on pre-2024 market conditions running in 2026 without retraining has drifted (Straive, 2026). Top-quartile manufacturers in Gartner's Global Supply Chain Top 25 achieve roughly 80% forecast accuracy at product mix level; most retail implementations enter at 10 to 15 percentage points below target coverage (Gartner Supply Chain Top 25, 2024; Netguru, 2025).

• The CFO is watching the data budget

  Grant Thornton's Q2 2026 CFO survey of 232 finance leaders found sharply declining execution confidence even as technology investment accelerated. The gap between ambition and delivery is growing, which means requests to remediate data infrastructure compete against AI tools that are easier to approve and faster to demo.

What the consumer actually did between 2022 and 2025

The training window matters. If the model ingested transaction data from 2019 through 2023, it captured a consumer who was stimulus-flush, discount-responsive in a relatively stable channel mix, and spending above income in a way that looked like confidence rather than strain. That consumer still exists in the wealthiest income decile. For everyone else, the behavior pattern shifted in ways that look subtle in aggregate data and severe at the SKU level. Credit card debt grew 5.5% in 2025 to $1.28 trillion (Yahoo Finance / Federal Reserve, 2025), while the personal savings rate dropped to 4.0% by Q1 2026, down from 6.2% in early 2024. Consumer spending growth has outpaced post-tax income growth for more than two years (Deloitte US Economic Forecast, 2026). The spending is real, but the fuel has changed. Debt is funding what income cannot, and when debt service pressure increases, the first adjustment is channel and brand, not total units. The behavioral fingerprint of that adjustment: 46% of shoppers modified their habits in response to rising costs in summer 2025, with 53% relying more on discount stores and 48% delaying major purchases (Trax, 2025). Over one-third of consumers traded down in some categories to spend in others they valued (Netguru, 2026). A model that expects this customer at a mid-tier channel at full price, based on 2022 purchase history, is reading the wrong person.

How trade-down specifically degrades a forecasting model's output

Trade-down does not reduce total transactions. It redirects them. That distinction is the one a model trained on historical channel-level data will miss. When a household that previously shopped a department store mid-tier moves to an off-price chain, two things happen to the model simultaneously. The training data still weights that household's historical basket toward the original channel. And the off-price retailer's volume looks like organic demand growth rather than diverted share. The result: the model over-forecasts the mid-tier channel and gives a clean signal on a category that is actually experiencing substitution, not growth. Circana's retail tracking confirmed the pattern in consumer electronics through H1 2025: unit demand for televisions rose 2% while sales revenue fell 3% (Circana, 2025). Consumers bought more but spent less per unit. A revenue-trained model sees softness. A unit-trained model sees growth. Neither sees the trade-down correctly without a price-tier feature that the original training data probably did not include. The Upside Consumer Spend Report (2025), drawing on over 10 billion transactions, documented what they called accelerating uncommitted behavior: consumers are more fragmented, more price-sensitive, and more value-driven than the prior year's data suggested. A model that treats historical loyalty coefficients as stable inputs is encoding a customer relationship that no longer exists for a significant portion of the base.

Where the forecast divergence shows up first

Model drift does not announce itself. It shows up three months later as a board question about margin or a planning meeting where the inventory position does not match the sell-through. The highest-risk divergence points are predictable:

• Mid-tier discretionary SKUs

  This is where trade-down hits hardest. Apparel, home goods, and non-essential consumables at moderate price points lose share to private label and off-price. A model without updated price elasticity features will hold its forecast flat while actual demand softens.

• Brand versus private label mix

  In summer 2025, 66% of consumers who changed their habits switched to cheaper or generic brands (Trax, 2025). If the model was trained when private label penetration was lower, it will underestimate the shift and produce an over-buy recommendation on branded SKUs.

• Channel allocation

  Increased traffic to value-oriented grocers, warehouse clubs, dollar stores, and off-price chains is documented through 2025 (Placer.ai, 2025). A model trained on pre-trade-down channel distribution will mis-route replenishment inventory.

• High-value SKU accuracy at the aggregate level

  Teams switching from MAPE to WMAPE as their primary evaluation metric frequently found their models were underperforming on high-value SKUs while appearing accurate in aggregate (Netguru, 2025). A model that looks fine in the weekly summary may be most wrong where it is most expensive to be wrong.

• BNPL-influenced demand patterns

  BNPL adoption among US adults grew from roughly 10% in 2021 to 15% by 2024, then held between 14% and 16% through 2025 (Statista, 2025; The Motley Fool, 2025). BNPL separates the purchase moment from the payment moment, and a model that uses payment or card data as a demand proxy will misread the timing of actual purchases.

The supplier reconciliation column nobody asked about

This is where the Monday morning moment comes back. The supplier reconciliation column that has been off by a rounding error for six weeks is almost always where the manual adjustments that make the report close are hiding. Someone on the analyst team found a discrepancy they could not explain, adjusted a number so the report would close on time, and moved on. Then did it again the following week. And the week after that. The report lands clean. The summary line matches expectations. The leader forwards it. The intention to ask about that column was real on Friday. By 8:47 Monday it is gone, absorbed by the three minutes between the inbox and the first call. A person carrying more than they should be carrying, running on whatever cognitive reserve remains after everything else has been served first, moves on. Asking feels like a luxury the calendar does not allow. I know that state from August 2026, when I was simultaneously ramping up and ramping down client work, saying yes to community commitments I had not fully evaluated, and setting intentions I kept almost keeping. The pressure of that much simultaneous movement means that the things requiring patience rather than action, the things that need you to sit with a question instead of answering one, get deferred until there is space. There is never space. So they stay deferred. The column is the signal. Supplier reconciliation data is often where a model producing a systematically off forecast gets quietly corrected by a planner who has learned, through experience, that the model cannot be trusted for that SKU cluster but has no formal mechanism to document that. The correction lives in their head and in the adjustment they make each week. The report looks fine. The model is getting worse. AI readiness in forecasting depends on whether the person approving the output knows what their team does to make the numbers look right. In most cases, they do not, and have not had the capacity to find out. That capacity problem is real. It also grew from something specific: the habit, practiced across hundreds of Monday mornings, of treating a clean summary line as permission to move on.

The one question to ask before the next planning cycle

The single question worth asking before the next planning cycle is this: what did the analyst change to make the supplier reconciliation column balance? One question. One column. Ask it before the call starts. If the answer is immediate and documented, you have a functioning feedback loop and the conversation can move to whether the model needs retraining. If the answer is 'let me check' or a pause that lasts longer than it should, you now know where the actual problem is. The model is the visible risk. The absence of visibility into how the model's output gets adjusted before it reaches you is the risk that compounds every week it goes unexamined. From there, the audit expands naturally. Pull the training window date on the current forecasting model, the accuracy on your top revenue SKUs measured at the SKU level rather than category, and the override log for the last 90 days with a comparison to actuals. If any of those three requests produces confusion about where to find the data, the audit has already told you what you need to know. But none of that happens without the first question, asked before the call takes the intention away.

The CFO lens: what softening consumer spending does to the data budget

Here is the second-order operational problem. The right response to a drifted forecasting model is to remediate the data infrastructure: retrain the model on current transactions, add price-tier and channel features, build an override audit loop. That costs money, headcount, and project bandwidth. At the same time, the CFO is watching consumer spending soften. Consumer confidence fell to its lowest level in more than a decade in early 2026, according to the Conference Board (CFO Dive, 2026). Consumer spending growth has outpaced post-tax income growth for more than two years, and Deloitte's 2026 economic forecast expects that gap to close from the spending side (Deloitte, 2026). When the CFO models that scenario, they see margin pressure and reach for the capital expenditure lever. Grant Thornton's Q2 2026 survey of 232 CFOs found sharply declining confidence in execution even as technology investment nominally accelerated. The gap between ambition and delivery is widening. In that environment, a request to remediate data infrastructure for a model that is 'mostly working' competes against AI tools that are easier to approve, faster to demo, and carry a visible ROI story. The data quality investment that would actually fix the forecasting problem gets deferred in favor of a new tool that will be trained on the same degraded data. Protiviti's 2026 survey of 1,540 global executives found that 39% of CFOs ranked infrastructure upgrades among their top investment priorities, and that outdated systems slow AI deployment and weaken forecasting accuracy (Protiviti, 2026). The recognition is there. The execution confidence is not. The merchandising or supply chain leader who wants budget to fix this problem needs to frame it the way a CFO processes risk: what is the inventory carrying cost of the next forecast miss, and how does that compare to the cost of retraining the model on current data? A three-point divergence in forecast accuracy at the SKU level translates directly to overstock write-downs or stockout-driven lost sales. That is a number. Put it in the room.

What could go wrong

• Retraining on current data without addressing channel and price-tier features

  Feeding 2024 and 2025 transaction data into a model that was not built to encode channel migration or price-tier substitution produces a model that is current but still structurally blind to the trade-down pattern. Retraining is necessary but not sufficient.

• Treating a clean summary report as a clean model

  Aggregate forecast accuracy can look adequate while the model is most wrong on the highest-value SKUs. Running accuracy metrics only at category or department level masks where the inventory risk actually sits.

• Manual overrides without a feedback loop

  If planners are correcting the model output and those corrections are never compared to actuals, the team is accumulating an unknown bias in the background. The model looks fine. The planners feel useful. The forecast is getting worse.

• Losing the data infrastructure budget to an easier AI story

  In a capital-constrained CFO environment, a new AI tool that demos well will almost always win budget over a data remediation project that requires months of work before it produces a visible output. The consequence is a better-looking tool running on a worse data foundation.

• Assuming consumer behavior will revert

  The trade-down pattern documented through 2024 and 2025 is structural. Consumer debt is at a record, the savings rate is down, and income growth is decelerating (Deloitte, 2026). A planning assumption that demand will return to pre-2024 patterns in the absence of significant debt relief or income acceleration is not supported by the data.

• The income bifurcation problem

  If a retailer's customer base has historically skewed middle income, the top-10% spending concentration documented by Moody's Analytics (Forbes, 2025) is not good news for aggregate demand forecasting. The wealthy segment does not offset the trade-down behavior of the middle; it distorts aggregate channel data in ways that can mislead a model that treats the customer base as uniform.

The J.Caresse point of view

The consumer debt story and the AI forecasting story share a mechanism. Something changed in the environment, the change accumulated quietly, and by the time it showed up visibly, the decisions that made it worse had already been made. The consumer did not announce that they were trading down. They just started appearing in a different store, buying a different brand, at a different price point. The model did not announce that it had drifted. It kept producing outputs that looked reasonable in aggregate while the inventory position degraded at the SKU level. Both problems persisted because the report looked fine. The person who could have caught this earlier approved it without asking. They approved it because they had three minutes, the summary line matched, and the intention to dig into that one column got absorbed by the call that started at 8:50. They approved it the week before for the same reason, and the week before that. A burnt-out leader who has been carrying more than their share long enough develops a habit of moving on without asking, and that habit starts to feel like efficiency. What it produces, over six weeks of Monday mornings, is a forecast model quietly corrected by a planner and a board question nobody can answer. The sentence worth keeping: a summary line matching your expectation tells you the report closed, and nothing more about whether the numbers are right. Say that before you forward the report. Ask one question about one column. That is the whole ask.

Key takeaways

Six findings to carry into the next planning meeting.

• The training data predates the behavioral shift

  A model trained through 2023 was built on a consumer who is no longer the dominant pattern. US household debt reached $18.8 trillion in Q4 2025 and the personal savings rate has declined sharply. That is a different demand signal than the one the model learned from.

• Trade-down is a channel and price-tier event, not a volume event

  Consumers are not buying less in aggregate. They are buying differently: different stores, different brands, different price points. A model without channel and price-tier features will misread substitution as decline and migration as growth.

• Aggregate accuracy metrics hide SKU-level risk

  Teams that switched to WMAPE found their models were underperforming on high-value SKUs while appearing accurate in aggregate. The most expensive forecast errors are the ones a MAPE-only evaluation will not surface.

• Override governance is where bias hides

  Planner overrides reduced forecast accuracy in roughly 40% of cases in retail and manufacturing engagements between 2023 and 2025, because there was no closed-loop comparison of override to actuals (Netguru, 2025). If the team overrides the model and no one audits the outcome, the bias accumulates without a name.

• The CFO will defer the data fix unless you frame the inventory cost

  In a capital-constrained environment, data remediation competes against tools that demo well. The way to get the budget is to translate forecast divergence into carrying cost: what did the last three-point miss cost in write-downs or lost sales? That is the number the CFO needs to hear.

• The Monday morning report is where AI readiness is decided

  Every week the report lands and gets forwarded without a question about data sourcing is a week the underlying problem compounds. The sentence worth saying at 8:47 AM: the summary line matching my expectation is not the same thing as the numbers being right.